Investigation of lifelong learning methods with elastic weight consolidation (EWC) for low-temperature ORC scroll expander modeling

This study applied a lifelong learning method, elastic weight consolidation (EWC), to enhance the memory stability and plasticity of deep learning models for organic Rankine cycle (ORC) expander modeling. We conducted multiple experiments using a low-temperature ORC power system with a scroll expander. Three deep-learning models were developed to predict the power output, exhaust temperature, and flow rate. Six task sequences consisting of combinations of three tasks were generated. Joint training (JT), fine-tuning (FT), and EWC were evaluated using these task sequences. The overall performance of the EWC and FT methods on model plasticity was close to that of the JT method across all three predictions (e.g., the largest MSE difference was only 61 for power prediction), but the performance of them on model memory stability was inferior to that of JT. Compared with the FT method, the EWC method showed good improvement, limited improvement, and no improvement in model memory stability for power, exhaust temperature, and flow rate prediction, respectively. In terms of MSE, the average performance increased by 18% for power prediction, but reduced by 76% for flow rate prediction. The effects of dataset sizes on the performance of lifelong learning were also evaluated. The performance of the EWC and FT methods stabilized when the dataset size reached approximately 300, which means that there is a threshold of dataset sizes for EWC practical application. The study results advance the application of deep learning modeling and offer insights into using EWC for scroll expander modeling.